Widely used vision-based datasets, such as DexYCB, often suffer from a vision-physics mismatch, exhibiting artifacts like interpenetration or missed contacts when replayed in simulators. To bridge this gap, we propose a physics-integrated refinement f...
Widely used vision-based datasets, such as DexYCB, often suffer from a vision-physics mismatch, exhibiting artifacts like interpenetration or missed contacts when replayed in simulators. To bridge this gap, we propose a physics-integrated refinement framework that transforms physically flawed demonstrations into dynamically executable trajectories. We formulate refinement as a physics-based trajectory optimization problem, optimizing only the hand's actuation parameters to induce natural object motion. To ensure efficiency, we introduce an adaptive, failure-driven optimization strategy that triggers refinement only when physical inconsistencies are detected. This strategy allows the optimizer to dynamically skip over physically consistent segments, concentrating computational resources on resolving complex contact failures. By optimizing spline-parameterized corrective deltas via CMA-ES, we ensure temporal smoothness and computational efficiency.
Extensive experiments demonstrate that our approach significantly outperforms state-of-the-art policy-based methods, achieving a computational speedup of approximately 4.2X while consistently attaining lower tracking errors in both object pose and hand pose. Our framework effectively recovers physically valid, contact-rich interactions from noisy vision data.